CNN-Based Image Classification

Intel Image Classification Dataset

Project Overview

This project implements a Convolutional Neural Network (CNN) for multi-class image classification using the Intel Image Classification Dataset.

The model classifies natural scene images into six categories:

• Buildings
• Forest
• Glacier
• Mountain
• Sea
• Street

This project covers data preprocessing, CNN model development, training, evaluation, and performance visualization.

---

Objective

To design, implement, and evaluate a CNN model for image classification using TensorFlow/Keras, including:

• Data preprocessing
• Data augmentation
• CNN architecture design
• Model training and validation
• Performance evaluation
• Prediction visualization

---

Dataset

• Dataset Name: Intel Image Classification
• Source: Kaggle
• Total Classes: 6
• Image Size Used: 150 × 150
• Dataset Split:
  • Training: 80%
  • Validation: 20%
  • Test: Separate test folder

---

Technologies Used

• Python
• TensorFlow / Keras
• NumPy
• Matplotlib
• Seaborn
• Scikit-learn
• Google Colab (GPU)

---

Data Preprocessing

The following preprocessing steps were applied:

• Resizing images to 150×150
• Pixel value normalization (0–255 → 0–1)
• Train-validation split (80-20)
• Data augmentation:
  • Random horizontal flip
  • Random rotation
  • Random zoom

Data augmentation improves model generalization and reduces overfitting.

---

CNN Architecture

The CNN model includes:

• 3 Convolutional Layers (ReLU activation)
• MaxPooling layers
• Flatten layer
• Fully Connected (Dense) layer
• Dropout (0.5) for regularization
• Output layer with Softmax activation (6 classes)

Model Flow:

Input Image
→ Rescaling
→ Data Augmentation
→ Conv2D + ReLU
→ MaxPooling
→ Conv2D + ReLU
→ MaxPooling
→ Conv2D + ReLU
→ MaxPooling
→ Flatten
→ Dense (128)
→ Dropout
→ Output (Softmax)

---

Training Details

• Optimizer: Adam
• Loss Function: Sparse Categorical Crossentropy
• Metrics: Accuracy
• Callbacks:
  • Early Stopping
  • Model Checkpoint

---

Model Evaluation

The model was evaluated using:

• Accuracy
• Precision
• Recall
• F1-score (per class)
• Confusion Matrix

Training and validation accuracy/loss curves were plotted to analyze overfitting.

---

Results

• Achieved strong classification performance across six classes.
• Confusion matrix analysis showed minor misclassifications between similar classes (e.g., Glacier vs Mountain).
• Data augmentation improved validation accuracy.

---

Prediction Visualization

Random test images were displayed with:

• Predicted Label
• Actual Label

This helped interpret model performance visually.

---

Future Improvements

• Use Transfer Learning (MobileNetV2 / ResNet50)
• Hyperparameter tuning
• Batch Normalization
• Learning rate scheduling
• Fine-tuning pretrained models

---

Conclusion

This project demonstrates the complete workflow of building a CNN from scratch for image classification.

It provides practical understanding of:

• Convolutional Neural Networks
• Model training and evaluation
• Overfitting control
• Performance analysis

---

Author

Fathima Salga
Deep Learning Enthusiast | Data Science Learner